The Theranos story gets a lot clearer when you look at it through the lens of regulatory compliance. Here was a diagnostics company that raised a ton of money before collapsing into a criminal conviction, but for anyone working in healthcare AI, the real lesson is the evidence gate they deliberately bypassed, not the salacious founder drama.
Theranos and the Evidence Gate
Eric Topol’s commentary on the Theranos collapse really nails it by focusing on the company’s wild claims, the total lack of independent review, and the fact that untested products made it to patients. They just sidestepped the entire evidence gate that’s supposed to exist for healthcare product development, a failure that directly led to the company’s criminal conviction.
The normal route for a new device claim is through the FDA’s CDRH, specifically the 510(k) pathway which builds trust through a documented process. Theranos had none of that, and its lack of any peer-reviewed clinical validation is what defines the entire scandal. The house of cards fell when FDA CDRH finally issued a warning letter in 2015, an action that forced the company to void every single test result and exposed the fraud FDA warning letter Theranos 2015.
The Expected Review Path for AI Health Tools
For most medical devices, including the majority of cardiac AI tools, the main path to market is the FDA 510(k) Clearance pathway where you have to show your product is substantially equivalent to something already on the market. The big headache for AI/ML is what happens when you retrain your model on new data, and the FDA’s Predetermined Change Control Plan (PCCP) is the answer to that. Without a PCCP in place, you’d be stuck submitting a brand new 510(k) every single time your adaptive AI learns something new, which is just unworkable.
When you’re doing due diligence, you need to ask about Good Machine Learning Practice (GMLP), which are the 10 guiding principles published jointly by the FDA, Health Canada, and the MHRA for building safe AI/ML devices. A company that hasn’t built its tech around GMLP is just accumulating regulatory debt that will have to be paid later. On top of that, you need a Quality Management System (QMS). For a CE mark, it has to be ISO 13485 compliant, and since the FDA’s Quality Management System Regulation (QMSR) formally adopted ISO 13485:2016, it’s now the foundational quality framework for U.S. medical devices as of February 2, 2026.
Recorded Failure Set: Olive AI and Pear Therapeutics
You can add Olive AI and Pear Therapeutics to the list of failures, and in each case, they acted like the evidence layer was optional. Olive AI’s marketing claims wildly outpaced its actual evidence, with zero publications to back its products Audit of Olive AI claims vs evidence. Pear Therapeutics presents a different flavor of failure: they actually got FDA clearance but still went bankrupt in April 2023 and sold off their assets by September 2024, showing that the marketing and the science simply didn’t align for a viable business. The common denominator is a disconnect between claims and real, peer-reviewed validation.
If you actually audit the marketing claims of AI health companies against their published evidence, you see a whole spectrum of behavior. At one extreme, you have Theranos, which just flat-out fabricated its data. Then you have companies like Olive AI and Pear Therapeutics, where the claims just didn’t have enough solid evidence to back them up. It all just proves you can’t take vendor assertions at face value and have to check them against data you can actually verify.
Positive Signals: FDA Clearance/Approval History and Published Outcomes
So what are the positive signals? As a buyer, one of the first things you can do is check a company’s FDA Clearance/Approval History yourself. Look at companies like HeartFlow, whose claims are backed by over 600 peer-reviewed publications on top of their FDA clearance and NICE approval. Or Big Health, with 33 papers and 8 solid Randomized Controlled Trials (RCTs). Hello Heart is another good one. Their claims of a 20.9 mmHg systolic reduction at 3 years are right there in a 2021 JAMA Network Open paper, and they’ve followed up with a 2025 Value in Health publication showing cost savings and reduced hospital use for heart failure patients, plus they’re working with the American College of Cardiology. These are the kinds of vendors who aren’t afraid of the evidence.
Beyond the big, formal trials, you have to look for Real-World Evidence (RWE). This is the data that comes from messy, real-life sources like EHRs, patient registries, and insurance claims. Using RWE makes your FDA submission stronger and also builds a much more convincing story for payers who want to see how a product actually performs out in the wild, across all kinds of different patients, not just in a controlled trial environment. A company that bothers to collect and publish RWE is showing you they’re serious about the evidence.
Compliance as the Gatekeeper
Think of regulatory compliance as the gatekeeper standing between a marketing claim and a real product. The spectacular collapse and criminal conviction of Theranos is what happens when a company decides to just hop the fence. It brings us back to Eric Topol’s core point: you absolutely need independent review before any of this stuff gets near a patient. It isn’t just paperwork.
For any health system safety officer looking at new AI tools, the takeaway is simple: put vendors with a solid FDA Clearance/Approval History and a deep list of peer-reviewed clinical validation at the top of your list. The cautionary tales of Olive AI and Pear Therapeutics show what happens when you don’t. A trustworthy AI platform is one that’s transparent with its evidence and follows the rules, because that compliance is what in the end keeps patients safe and protects the health system itself FDA guidance on medical device regulatory compliance.
Frequently Asked Questions
What is the primary regulatory pathway for most cardiac AI products?
The FDA 510(k) Clearance pathway is the primary route for most cardiac AI products. This pathway demonstrates substantial equivalence to a predicate device. It is the expected route for a claim to be validated before reaching patients.
What is the significance of Peer-Reviewed Clinical Validation for AI health tools?
Peer-Reviewed Clinical Validation is a critical trust indicator. Its absence defined the Theranos case, where claims bypassed expected validation. Companies like HeartFlow, Big Health, and Hello Heart demonstrate strong evidence bases through extensive peer-reviewed publications.
What is a Predetermined Change Control Plan (PCCP) and why is it important for adaptive AI?
A PCCP is an FDA framework that allows AI/ML devices to make predefined modifications without new premarket submissions. This is critical for adaptive cardiac AI. Without a PCCP, every AI model retraining on new data would require a new 510(k) submission.
What quality standards should vendors of AI health tools comply with?
Vendors should comply with Good Machine Learning Practice (GMLP) principles, which ensure safe and effective AI/ML medical devices. Additionally, a Quality Management System (QMS) compliant with ISO 13485 is required for CE marking and is the core quality framework for medical devices in the U.S. as of February 2, 2026.
